COMPUTATIONALLY EFFICIENT PARTICLE FILTERING USING ADAPTIVE TECHNIQUES

We propose a computationally efficient particle filtering al go- rithm that adaptively chooses between the sequential impor- tance resampling (SIR) particle filter and the unscented par - ticle filter (UPF). The technique is based on the use of the Kullback-Leibler distance (KLD) sampling and the choice of either of the algorithms is governed by the error in estimati on. The SIR particle filter most opted among the variations of the particle filter because of the choice of the transitional pri or as the importance density and easy evaluation of weights. How- ever, it can be inefficient for highly non-linear dynamic sys - tems . In contrast, the UPF which uses the scaled unscented transform performs better than the SIR but is computation- ally more expensive. The proposed algorithm couples the easy evaluation of the weights and the faster sampling capa- bilities of the SIR filter with the improved accuracy of the UPF. We apply the technique to a scalar estimation problem and demonstrate through simulations that the new algorithm is more accurate than the SIR particle filter and is faster tha n the UPF for systems characterized by highly non-linear mea- surement models. of particle filtering algorithms so that they can be applied t o real-time systems. Different choices of importance density (2, 3, 4) and varying modification in the resampling stage (5, 6, 7) have yielded various enhancements to the generic particle filter. Our work is based on the SIR particle filter and the UPF. The UPF performs better than the SIR for highly non-linear scenarios; however, this increase in estimation accuracy i s traded-off with a higher computational complexity. Our ob- jective is to propose a new particle filtering approach for highly non-linear dynamic scenarios that maintains the ac- curacy of the UPF but is closer in computational complexity to the SIR. Towards this goal, we first analyze the perfor- mance and computational complexity of the SIR and the UPF in Section 2. We then develop a new system for sequen- tial Bayesian filtering in which the SIR particle filtering an d the UPF algorithms are adaptively configured based on the Kullback-Leibler distance (KLD) sampling technique. The implementation is aimed at achieving a computationally ef- ficient system with reduced estimation error. The detailed description of the technique is provided in Section 3. In order to demonstrate the improved performance of the new algorithm, we test it using a one-dimensional (1-D) scalar es- timation problem and provide the simulation results in section 4.